One interesting thing is that using reservoir sampling "Algorithm L"[0], you can actually get an unbiased sample with overhead proportional to the logarithm of CPU time. That would require more state and work in the server to aggregate, though. On the other hand, moving overhead from the process to the aggregation server is useful if you want to run profiling in production.

[0] https://en.wikipedia.org/wiki/Reservoir_sampling#An_optimal_...